A Simple yet Powerful Deep Active Learning with Snapshots Ensembles

Official Implementation of the paper "A Simple yet Powerful Deep Active Learning with Snapshots Ensembles"

---

1. Introduction

---

2. How to Use

2-1. How to Run

Run experiments on CIFAR-10 dataset with ResNet-18. Specify an acquisition function with --query_type and acquisition size with --query_size argument.

python scripts/train_snapshot.py -f configs/cifar10_resnet18.json --query_type vr

python scripts/train_ensemble.py -f configs/cifar10_resnet18.json --query_type vr

python scripts/train_sameinit.py -f configs/cifar10_resnet18.json --query_type vr

Run experiments on CIFAR-100 dataset with ResNet-18. Specify an acquisition function with --query_type and acquisition size with --query_size argument.

python scripts/train_snapshot.py -f configs/cifar100_resnet18.json --query_type vr

python scripts/train_ensemble.py -f configs/cifar100_resnet18.json --query_type vr

python scripts/train_sameinit.py -f configs/cifar100_resnet18.json --query_type vr

Run experiments on Tiny-ImageNet-200 dataset with ResNet-50. Specify an acquisition function with --query_type and acquisition size with --query_size argument.

python scripts/train_snapshot.py -f configs/tiny_resnet50.json --query_type vr

python scripts/train_ensemble.py -f configs/tiny_resnet50.json --query_type vr

python scripts/train_sameinit.py -f configs/tiny_resnet50.json --query_type vr

2-1-1. Download Datasets

CIFAR10 and CIFAR100 dataset will be downloaded using torchvision and saved to a directory provided as --dataset_path.

However, for Tiny Imagenet, you will need to download it first, unzip it, and specify a unzipped directory as --dataset_path.

Our code will automatically create a file structure for the validation set.

Please edit config files according to your local setting before running the scripts.

2-1-2. Edit Configs

Configuration files are in ./cifar/configs and ./tiny-imagenet/configs folder. You can directly edit these files for different hyperparameters.

---